[Paper Review] MotionDiffuse: Text-Driven Human Motion Generation with Diffusion Model

Mingyuan Zhang, Zhongang Cai+5 more|arXiv (Cornell University)|Aug 31, 2022

Human Motion and Animation110 citations

TL;DR

MotionDiffuse uses a diffusion-model framework with a cross-modality transformer to generate diverse, controllable text-driven human motions, including body-part and time-varied control.

ABSTRACT

Human motion modeling is important for many modern graphics applications, which typically require professional skills. In order to remove the skill barriers for laymen, recent motion generation methods can directly generate human motions conditioned on natural languages. However, it remains challenging to achieve diverse and fine-grained motion generation with various text inputs. To address this problem, we propose MotionDiffuse, the first diffusion model-based text-driven motion generation framework, which demonstrates several desired properties over existing methods. 1) Probabilistic Mapping. Instead of a deterministic language-motion mapping, MotionDiffuse generates motions through a series of denoising steps in which variations are injected. 2) Realistic Synthesis. MotionDiffuse excels at modeling complicated data distribution and generating vivid motion sequences. 3) Multi-Level Manipulation. MotionDiffuse responds to fine-grained instructions on body parts, and arbitrary-length motion synthesis with time-varied text prompts. Our experiments show MotionDiffuse outperforms existing SoTA methods by convincing margins on text-driven motion generation and action-conditioned motion generation. A qualitative analysis further demonstrates MotionDiffuse's controllability for comprehensive motion generation. Homepage: https://mingyuan-zhang.github.io/projects/MotionDiffuse.html

Motivation & Objective

• Lower the barrier to create realistic human motions from natural language descriptions.
• Introduce a probabilistic, diffusion-based approach to text-to-motion generation for higher diversity.
• Enable multi-level manipulation including body-part-level control and arbitrary-length motion synthesis.
• Demonstrate state-of-the-art performance on text-driven and action-conditioned motion generation tasks.

Proposed method

• Use a diffusion model (DDPM) to generate motion sequences conditioned on text descriptions.
• Introduce a Cross-Modality Linear Transformer with a text encoder and a motion decoder to handle variable-length sequences.
• Incorporate Linear Self-Attention (Efficient Attention) and Linear Cross-Attention to fuse text into motion generation.
• Apply a Stylization Block to inject text and time (t) information into each denoising step.
• Implement body-part independent control via noise interpolation across partitioned body parts with smoothing corrections.
• Enable time-varied control by denoising multiple intervals and interpolating noises with a correction term.
• Train by optimizing a single loss that predicts the noise term psilon_theta in the diffusion process.

Experimental results

Research questions

• RQ1 Can diffusion models produce diverse, high-fidelity motions from natural language prompts?
• RQ2 Does a cross-modality transformer effectively fuse text into motion generation for variable-length sequences?
• RQ3 Can fine-grained, body-part-level and time-varying prompts be controlled during motion synthesis without sacrificing quality?
• RQ4 How does MotionDiffuse perform on text-driven and action-conditioned motion generation benchmarks compared to prior state-of-the-art methods?

Key findings

• MotionDiffuse achieves significant improvements over prior state-of-the-art methods on text-driven motion generation and action-conditioned motion generation.
• The framework demonstrates high-fidelity, diverse motion synthesis driven by natural language prompts.
• Multi-level manipulation allows body-part-level control and time-varied sequence generation without additional training cost.
• Qualitative analysis shows MotionDiffuse’s controllability and ability to handle complex, long motion sequences.
• Experiments on multiple datasets (e.g., HumanML3D, KIT-ML, HumanAct12, UESTC) illustrate broad applicability and superiority over existing approaches.